{
  "schema_version": 1,
  "builder_plan": {
    "scenario": "A narrow, repeated product task may not need your largest model.",
    "change": "Check whether a released student model fits the task before generating training data.",
    "baseline": "Your current model and an off-the-shelf small model on the same held-out tasks.",
    "steps": [
      "Locate the paper's implementation and check its license, model support, data needs, and hardware requirements. Stop here if you cannot obtain the required artifacts.",
      "Save a small set of representative product tasks and expected outcomes. Keep evaluation tasks out of any training or tuning, and choose quality and cost limits before running.",
      "Run the baseline and the proposed change on identical inputs. Keep other settings fixed and repeat paired runs to expose variation.",
      "Save per-task outputs, failures, timings, and all usage costs. Review regressions before expanding the test."
    ],
    "measure": "Student quality, serving cost, teacher calls, training cost, and failures on unusual inputs.",
    "decision": "Keep the change only if it meets your preselected quality and latency limits and improves the metric you care about after setup and operating costs. Otherwise retain the baseline. This is an exploration plan, not a validated recommendation."
  },
  "builder_review": null,
  "arxiv_id": "2610.11529",
  "title": "ReTeach: Building a Self-Teacher through Multi-Round Reflection and Retry",
  "summary": "It replaces self-distillation pipelines that depend on reference answers, external diagnostic feedback, or persistent cross-example memory by building the self-teacher through multi-round reflection and retry using only self-generated attempts and outcome-level verification. It costs multiple teacher rollouts per failed example plus on-policy token-level distillation compute, which raises training time substantially over single-pass GRPO while inference cost stays unchanged. It can fail when the model cannot correct itself within the retry budget (unresolved examples add noise), when outcome-level verification is unavailable or unreliable for the target task, and the 1.39-point average gain may not justify the added training cost on a limited budget.",
  "rationale": "The method is reproducible on a single GPU since it requires no separate teacher model and keeps single-pass inference, but the modest average gain and heavy multi-rollout training cost mean the reader should validate it on their own workload before adopting it over a plain GRPO baseline.",
  "technique": "Reflection-based self-distillation with retry",
  "family": "destilacao",
  "recommendation": "testar",
  "published": "2026-10-08",
  "as_of": "2026-10-18",
  "editorial_status": "indexed",
  "source_url": "https://arxiv.org/abs/2610.11529",
  "independent_implementations": 1,
  "total_implementations": 2,
  "stars": 146,
  "citations": 0,
  "claims": [
    {
      "claim_id": "claim-01",
      "statement": "It replaces self-distillation pipelines that depend on reference answers, external diagnostic feedback, or persistent cross-example memory by building the self-teacher through multi-round reflection and retry using only self-generated attempts and outcome-level verification. It costs multiple teacher rollouts per failed example plus on-policy token-level distillation compute, which raises training time substantially over single-pass GRPO while inference cost stays unchanged. It can fail when the model cannot correct itself within the retry budget (unresolved examples add noise), when outcome-level verification is unavailable or unreliable for the target task, and the 1.39-point average gain may not justify the added training cost on a limited budget.",
      "basis": "inferred",
      "result": "",
      "baseline": "",
      "conditions": "",
      "source_url": "https://arxiv.org/abs/2610.11529",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "claim_id": "claim-02",
      "statement": "ReTeach improves average accuracy over GRPO by 1.39 percentage points across six benchmarks spanning mathematical reasoning, science QA, and tool use.",
      "basis": "inferred",
      "result": "",
      "baseline": "",
      "conditions": "",
      "source_url": "https://arxiv.org/abs/2610.11529",
      "source_page": null,
      "source_excerpt": ""
    }
  ],
  "exposure_map": [
    {
      "dimension": "quality",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "compute",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "latency",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "operations",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "compatibility",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "security",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "data_and_training",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "reproducibility",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    }
  ],
  "risks": [],
  "minimum_test": [],
  "open_questions": [],
  "independent_tests": [],
  "lab_reproductions": [],
  "report_available": false,
  "validation_tier": "unknown",
  "evidence_tier": "unknown",
  "infrastructure_basis": "unknown",
  "equations": [],
  "equations_status": "not_fetched",
  "core_kind": "",
  "equations_fetched_at": "",
  "equations_source_url": "",
  "equations_source_sha256": "",
  "exposures_reviewed_at": "",
  "exposures_source_sha256": ""
}